Add real-path grader smoke tests for training-based tasks

[GTC2333]

Problem

The BCI and Sleep Oracle modes skip model import, data preprocessing, training, and inference, and directly emit predefined chance-level scores. Passing Oracle validation therefore confirms only that the scorer can emit valid JSON, not that the real grader path works.

Trigger scenario

A change introduces a regression in preprocessing, input units, epoch alignment, model loading, Docker mounts, training, inference, or prediction ordering. Oracle validation still passes, while official scoring fails or produces incorrect results.

Impact

• Validation gives false confidence in grader correctness.
• Regressions may only appear after an expensive full benchmark run.
• The current Oracle path cannot protect the real evaluation pipeline.

Expected fix

Keep the current fast Oracle, but add a small evaluator-only golden fixture that exercises the real path:

• run actual preprocessing and model import;
• perform a minimal training and inference step;
• compute real metrics;
• cover Docker mounts and submission isolation;
• remain small enough for CI and task validation.

This does not require running the full nine-subject BCI evaluation or the full Sleep-EDF training workload.